Projects

### Overview: This repository provides information and resources related to open source speech datasets, offering diverse and high-quality speech data across various domains such as conversational, academic, political, and more. These datasets are widely used for tasks like automatic speech recognition (ASR), speaker identification, emotion recognition, and other speech processing applications. ### Key features: - Casual Conversations - Common Voice ### Tech stack: Unknown ### Why it matters: These datasets are crucial for researchers and developers looking to improve speech processing algorithms and applications. They provide a rich resource for training models and testing their performance across different domains, making them essential for advancing the field of speech technology.

Overview: Cluaiz is a Rust-based local AI engine that supports running GGUF and ONNX models for various tasks such as chat, vision, speech, and embeddings. It automatically places models based on available hardware and supports OpenAI-compatible APIs. Key features: - Runs GGUF models through a fork of llama.cpp - Supports ONNX models through ONNX Runtime - A placement planner that decides where models run, context size, and KV cache type - Supports MoE models larger than GPU and RAM - Provides APIs for existing clients and a CEP protocol - Tools for agents, including skills, plugins, and MCP servers Tech stack: Rust Why it matters: Cluaiz simplifies the deployment of AI models by providing a unified solution that can run on local hardware, ensuring optimal performance and resource utilization. Repository: cluaiz/cluaiz GitHub stars: 28

Overview: This project provides a pure C++ and HIP inference solution for OpenAI's `gpt-oss-20b` and `gpt-oss-120b` models, utilizing AMD GPUs without relying on external libraries such as rocBLAS, RCCL, or MPI. The model achieves high throughput and fidelity by writing every kernel, collective, and tokenizer from scratch. Key features: - Pure C++ and HIP inference - No external dependencies (except HIP runtime) - High throughput: 84,451 tokens/s on 20B model, 40,271 on 120B model - High fidelity: METEOR and BERTScore scores above 0.3 and 0.9 Tech stack: Pure C++ and HIP, HIP runtime Why it matters: This project offers a lightweight and efficient alternative to large model dependencies, focusing on performance and fidelity without compromising on the model's capabilities.

Overview: This project involves scraping fraudulent emails from an anti-fraud website, deduplicating the dataset using MinHash and LSH, and fine-tuning a GPT-2 model on the deduplicated dataset. Key features: - Web scraping of scam emails - Deduplication using MinHash and LSH - Fine-tuning GPT-2 on a scam dataset Tech stack: Python Why it matters: This project aims to create a new dataset of fraudulent emails that can be used to advance research on intelligent email assistants.

### Overview: This is a patched native-Windows build of vLLM, featuring three specific fixes for Windows environments: CPU-relay for Gloo, Qwen3 reasoning parser fix, and wildcard model name support. It includes prebuilt wheels for Ampere/Ada and Blackwell GPUs, and a matching Qwen3.6-enhanced chat template. ### Key features: - **CPU-relay for Gloo collectives** - **Qwen3 reasoning parser fix** - **Wildcard model name support** ### Tech stack: - Python ### Why it matters: This patched build of vLLM is optimized for Windows environments, ensuring smooth performance and compatibility with Windows-specific tools and libraries.

Overview: This project is designed to automatically generate an audio-structured corpus for Automatic Speech Recognition (ASR) systems using pseudo-labeling techniques. Key features: - Integrates with YouTube or remote servers to collect content. - Supports direct links to videos or audio files. - Collects metadata including loudness, language detection, audio language detection, and SRMR ratio. - Provides options to export labeled data using whisper, wav2vec2, or NeMo as an ASR backend. - Supports multiple ASR models such as Whisper, Wav2Vec2, and NeMo. Tech stack: Python Why it matters: This tool simplifies the process of creating a corpus for ASR systems, making it easier for developers to test and improve their models without manual annotation.

Overview: This project records microphone audio, transcribes it locally with OpenAI Whisper, and copies the result to the clipboard. Key features: - Speech-to-text functionality - Integration with GNOME Wayland - Local transcription using Python and Whisper - Automatic paste to clipboard Tech stack: Python, Whisper, ffmpeg Why it matters: This project provides a practical solution for speech-to-text functionality on Linux, making it useful for developers and users who need to quickly transcribe audio recordings.

### Overview: This project introduces a speech recognition model for the Uyghur language, utilizing a transformer-based approach. The model is trained and tested using a dataset and pretrained model provided. ### Key Features: - **Training:** Uses Cross Entropy loss and CTC loss for training. - **Recognition:** Utilizes a pretrained model for recognition, with detailed output showing model performance metrics. ### Tech Stack: - Python ### Why It Matters: This project offers a practical solution for speech recognition in the Uyghur language, contributing to the development of language technology in the region.

### Overview: LawBotics v2 is an advanced legal technology platform that combines artificial intelligence with legal expertise to streamline contract review processes. The platform utilizes fine-tuned language models trained on the CUAD dataset to identify and extract key contract clauses, helping legal professionals save time and reduce errors in contract analysis. ### Key Features: - **AI-Powered Contract Analysis**: Automated clause identification and extraction from legal contracts. - **Multi-Format Support**: Processes contracts in PDF and text formats. - **Real-time Analysis**: Instant contract review with immediate results. - **Clause Categorization**: Identifies 41+ different types of legal clauses. - **Interactive UI**: Modern, responsive web interface built with Next.js 15. ### Tech Stack: TypeScript ### Why it Matters: LawBotics v2 leverages advanced AI and machine learning models to automate legal document review and clause extraction, making it an essential tool for legal professionals looking to streamline their contract analysis processes. Repository: hasnaintypes/lawbotics Primary language: TypeScript GitHub stars: 28 GitHub description: LawBotics v2 is an AI-powered legal contract analysis platform that combines machine learning with modern web technologies to automate legal document review and clause extraction.

### Overview: Vectorizer is a high-performance, in-memory vector database and search engine written in Rust, designed for semantic search and AI-driven applications. It ships as a Cargo workspace with binary RPC and HTTP transports, a React dashboard, and native SDKs for Rust, Python, TypeScript, Go, and C#. The database is optimized for speed, with SIMD acceleration and Metal GPU support, and it offers 4-5x faster search performance than Qdrant. It supports various storage formats, including `.vecdb` unified format, memory-mapped storage, and product and scalar quantization for efficient storage and retrieval. ### Key Features: - **Transport & API**: - **VectorizerRPC** (default, port `15503`) — binary MessagePack over TCP, multiplexed connection pool. - **REST API** (port `15002`) — universal HTTP fallback, powers the dashboard and any caller that doesn’t speak raw TCP. - **gRPC** — Qdrant-compatible service. - **GraphQL** — full REST parity with async-graphql + GraphiQL playground. - **MCP** — 31 focused tools for AI model integration (Cursor, Claude Desktop, etc.). - **UMICP Protocol** — native JSON types + tool discovery endpoint. - **Performance**: - **SIMD acceleration** — AVX2-optimized vector ops with runtime CPU detection (5-10x faster). - **Metal GPU** — macOS Apple Silicon via `hive-gpu

Overview: h-codex is a semantic code search tool designed for intelligent, cross-repo context retrieval, utilizing advanced techniques like AST-based chunking, embedding & semantic search, and efficient vector database storage. Key features: - AST-Based Chunking: Optimal code parsing using Abstract Syntax Trees. - Embedding & Semantic Search: Leveraging OpenAI's `text-embedding-3-small` model. - Vector Database: PostgreSQL with pgvector extension for efficient similarity search. - Multi-Language Support: TypeScript, JavaScript, and extensible for other languages. - Multi-Project Support: Index and search multiple projects. - MCP Integration: Seamlessly connects with AI coding assistants through Model Context Protocol. Tech stack: TypeScript Why it matters: h-codex offers advanced semantic code search capabilities, making it an invaluable tool for developers looking to enhance their productivity and understanding of code across projects. Repository: hpbyte/h-codex GitHub stars: 28 GitHub description: A semantic code search tool for intelligent, cross-repo context retrieval.

### Overview: This project presents a novel symmetric parameter-efficient fine-tuning framework for multimodal semantic segmentation, specifically addressing the challenges of high computational costs and limited transferability in RGB-based pre-trained models. The method, "Keep the Balance," aims to enhance the capabilities of pre-trained models to both RGB and X modalities by utilizing modality-aware prompting and adaptation, and dynamic sparse cross-modality fusion. ### Key features: - **Symmetric Parameter-Efficient Fine-Tuning Framework**: The method enables simultaneous adaptation of a powerful pre-trained model to both RGB and X modalities. - **Modality-Aware Prompting and Adaptation Scheme**: This scheme enhances the model's ability to learn from a diverse set of modalities. - **Dynamic Sparse Cross-Modality Fusion Module**: This module facilitates effective and efficient cross-modality fusion, mitigating the noise introduced by the attention mechanism. - **Self-Teaching Training Strategy**: The method leverages accurately predicted dual-modality results to self-teach single-modality outcomes, further strengthening the model. ### Tech stack: Python ### Why it matters: This framework addresses the limitations of traditional RGB-based segmentation methods by leveraging additional modalities effectively, thereby improving computational efficiency and broadening the model's transferability across different modalities. The method's ability to enhance the capabilities of pre-trained models to both RGB and X modalities makes it particularly valuable in applications requiring high accuracy and low computational cost.

Overview: SwiftLet is a lightweight Python framework designed for local execution of open-source Large Language Models (LLMs) using safetensors. Key features: - Local execution of supported LLMs - Minimal dependencies, easy to understand - Clean architecture for adding more models - Designed for research, prototyping, and educational use - Open for contributions and experiments Tech stack: Python Why it matters: SwiftLet provides a simple and educational way to run and experiment with various LLMs locally, making it ideal for learning and research purposes.

Overview: This application converts text into images, suitable for media sharing on platforms like WhatsApp or social media. Key features: - Converts text to images - Web interface for generating featured images Tech stack: Go Why it matters: This tool simplifies the process of creating featured images for media sharing, making it easier to post code snippets or other text-based content on social media platforms. Repository: johnpili/go-text-to-image Primary language: Go GitHub stars: 28

### Overview: Voice100 is a lightweight neural text-to-speech (TTS) and automatic speech recognition (ASR) model designed for efficient inference on various devices. It is particularly suitable for low-cost applications, such as smartphones and small embedded systems. ### Key Features: - **Tiny Models**: Voice100 is built with tiny models that do not require large datasets, ensuring compatibility with normal PCs, Raspberry Pi, and smartphones. - **CNN Dependencies**: The models are entirely dependent on CNNs, making them highly efficient and cost-effective. - **Inference Efficiency**: Inference is significantly faster compared to models that require complex autoregressive structures, making it ideal for real-time applications. ### Tech Stack: - **Python**: The primary programming language used for developing Voice100. ### Why It Matters: - **Efficiency**: Voice100's small size and reliance on CNNs make it highly efficient, suitable for running on low-power devices. - **Cost-Effectiveness**: The models are designed to be small, making them cost-effective for a wide range of applications, including those on smartphones and small devices. ### Repository: kaiidams/voice100 - **GitHub Stars**: 28 - **GitHub Description**: Voice100 includes neural TTS/ASR models. Inference of Voice100 is low cost as its models are tiny and only depend on CNN without autoregression. ### Sample Synthesis: - **English Sample Synthesis 1**: "beginnings are apt to be determinative and

### Overview: VideoSense is an AI agent designed to answer questions about your video library. It watches your videos, reasons about what it sees, and provides proof through clips and charts. The agent is built on Python and uses FastAPI for its backend, along with tools like Gemini 2.5, PostgreSQL, and Cloud Run for storage and processing. It's designed to be reliable and transparent, providing clear answers with detailed explanations and cost information. ### Key Features: - **Video Analysis**: Watches videos and reasons about what it sees. - **Answer Proofs**: Provides clips and charts to prove the answer. - **Interactive Demo**: A live demo showing a session and the answer. - **Detailed Explanation**: Answers are accompanied by timecodes, tool steps, and cost information. ### Tech Stack: - Python - FastAPI - Gemini 2.5 - PostgreSQL - Cloud Run - BigQuery ### Why It Matters: VideoSense offers a straightforward and reliable way to answer questions about your video library, providing proof and detailed explanations. It's built to be trusted and transparent, ensuring accurate and honest answers.

Overview: Pi-Grok-CLI integrates your X Premium or SuperGrok subscription into the Pi environment, providing a streamlined toolset for image generation and usage tracking. Key features: - OAuth login for seamless integration. - Support for multiple accounts and automatic quota management. - Detailed usage tracking and reset times. - Native image input capability. - Image generation directly or through the `image_gen` tool. Tech stack: TypeScript Why it matters: This tool simplifies the integration of premium Grok services into the Pi environment, making it easier to manage subscriptions and usage data efficiently. Repository: kenryu42/pi-grok-cli Primary language: TypeScript GitHub stars: 28

### Overview: This repository documents the patches needed to run NVFP4 quantized models at full speed on NVIDIA GeForce RTX 5090, 5080, 5070 Ti, and RTX PRO 6000 GPUs (SM120). It explains the critical insights into the unique architecture of SM120 and the necessary patches to make vLLM + FlashInfer + CUTLASS run correctly on these GPUs. ### Key features: - NVFP4 inference on Blackwell GeForce / RTX PRO 6000 - SM120 patches for vLLM + FlashInfer + CUTLASS - 175 tok/s on Qwen3.6-35B MoE on RTX 5090, 5080, 5070 Ti, and RTX PRO 6000 - SM120-specific GEMM and attention patches - Infrastructure patches for JIT compilation and runtime quantization ### Tech stack: Python ### Why it matters: This repository provides the critical patches needed to ensure that NVFP4 models run efficiently on the unique architecture of SM120 GPUs, ensuring optimal performance and compatibility with the latest cutting-edge models and frameworks.

### Overview: REPOCOD is a comprehensive code generation benchmark designed for evaluating language models in software development. It includes 980 problems from 11 popular real-world projects, focusing on tasks requiring file-level or repository-level context. REPOCOD features a long average canonical solution length and high average cyclomatic complexity, making it a valuable resource for assessing the capabilities of language models in real-world software development. ### Key Features: - 980 problems from 11 popular real-world projects - 58% of tasks requiring file-level or repository-level context - Longest average canonical solution length (331.6 tokens) - Highest average cyclomatic complexity (9.00) - 313.5 developer-written test cases per task for better correctness evaluation - Evaluation on ten LLMs, none achieving more than 30 pass@1 on REPOCOD ### Tech Stack: REPOCOD is built using Python, leveraging the Hugging Face dataset for tasks and evaluation. ### Why it Matters: REPOCOD highlights the necessity of developing stronger language models that can effectively assist developers in real-world software development, particularly in tasks requiring repository-level context information.

### Overview: Echo-XI is a Python-based project that utilizes the ElevenLabs TTS service to convert speech to text and vice versa. It offers a user-friendly interface and supports multiple speech recognition and text-to-speech providers. ### Key Features: - **Speech to Text and Text to Speech**: Converts audio to text and vice versa using ElevenLabs TTS. - **GUI and CLI**: Provides a graphical user interface and a command-line interface for easy access. - **Sensitive Details**: Stores API keys in the system keyring for security. - **Translation**: Automatically translates speech to the user's preferred language. - **Multiple Providers**: Supports Vosk, Azure Speech Recognition, and Whisper (both locally and through OpenAI's API). - **OBS Synchronization**: Synchronizes detected text with OBS text sources. - **Translation Services**: Uses DeepL for supported languages and Google Translate for others. - **Supports Multiple Languages**: Compatible with various languages through ElevenLabs' multilingual model. ### Tech Stack: - Python ### Why it Matters: Echo-XI simplifies the process of converting speech to text and text to speech, making it accessible and user-friendly. Its integration with multiple speech recognition and text-to-speech providers ensures versatility and reliability. The ability to translate speech into the user's preferred language and synchronize with OBS adds value to the project, making it a valuable tool for developers and enthusiasts alike. ### Repository: - Repository: `https://github.com/lugia19/Echo-XI` - GitHub Stars: 28 -